Open source · early access

Reliable memory for AI agents

Confidence computed from corroboration and conflict, not guessed by an LLM — and never hidden when sources disagree.

Open source Air-gapped by design EU AI Act aligned Energy sector focus Early access — testers wanted
The problem

Most agent memory quietly lies to you

Not maliciously — it just wasn't built to tell you when it's unsure, or when two things it was told don't agree.

1

Confidence, guessed

Ask an LLM how sure it is and you get a plausible-sounding number, not a computed one. Nothing to audit, nothing to reproduce.

2

Disagreement, hidden

When sources conflict, most memory systems quietly average or overwrite. The disagreement never surfaces — until it matters.

3

Context, not memory

Vector search returns what's similar, not what's true, corroborated, or contested. And context windows forget.

Why TekmerDB

Six ideas, one engine

Mechanical confidence

Confidence is computed from corroboration and conflict math, not guessed by an LLM — the same inputs always produce the same score.

Conflicts are signal, not noise

When two sources disagree, TekmerDB flags it and keeps both claims visible — instead of silently picking a winner.

Sources carry weight

Every claim ties to a source, and every source has an effective weight that rises with corroboration and falls when it triggers a conflict.

Reliable memory for agents

Every claim your agent ingests becomes a PFO — a Piece of Fact or Opinion — structured and persistent instead of ephemeral context.

Air-gapped & built for compliance

Runs fully offline and locks every instance to one EU AI Act risk category upfront — starting in the energy sector.

Domain scope keeps signal clean

Locking an engine to one EU AI Act risk category means claims are only ever compared within the same regulatory context.

How it works

From insert to answer

1

Insert a claim

Every insert is tied to a source. TekmerDB records it as a PFO — nothing about its confidence is decided yet.

2

Catch the conflict

Both claims stay. Confidence drops on the less-corroborated one — computed from the disagreement, not chosen by a model.

3

Query with confidence

Nothing gets silently dropped. Every claim returns with its confidence and, if relevant, exactly what it conflicts with.

Who it's for

Built for two kinds of teams

Agent builders

Give your agent persistent, queryable memory that doesn't quietly average away disagreement. Drop it in wherever you'd reach for a vector store.

Regulated teams

Air-gapped and scoped to a single EU AI Act risk category per instance — built for environments where you have to explain a number, not just produce one.

Help us break it

TekmerDB is early. Clone it, run it against real data, tell us what's wrong — feedback shapes what gets built next.